Sparse Inducing Points in Deep Gaussian Processes: Enhancing Modeling with Denoising Diffusion Variational Inference
Published in Proceedings of the 41st International Conference on Machine Learning (ICML 2024), 2024
The true posterior over Deep GP inducing variables is generally non-Gaussian, but mean-field Gaussian VI is the standard approximation. DDVI represents the variational posterior implicitly as the terminal state of a reverse-time variance-preserving SDE driven by a learned score network, jointly trained with a denoising score-matching regulariser. This sidesteps the mean-field restriction and gives an explicit, optimisable lower bound on the marginal likelihood.
Recommended citation: Xu, J., Zeng, D. and Paisley, J. (2024). Sparse Inducing Points in Deep Gaussian Processes: Enhancing Modeling with Denoising Diffusion Variational Inference. ICML 2024, PMLR 235, pp.55490-55500. (Oral)
Download Paper